colexec: fix range stats operator with the selection vector

[yuzefovich]

Previously, if the input batch had a selection vector set, we could get an index out of bounds internal error because we would iterate up to the length of the selection vector which can exceed the length of the batch.

Note that in practice this should be very rare to hit - generally speaking, the selection vector has the same length as all of the vectors in the batch. I couldn't reproduce the problem when using demo, so there is no regression test. Introduction of unit tests that use powerful RunTests test harness (that might have caught this bug) is left as a TODO.

Fixes: #93806.
Epic: CRDB-22701

Release note (bug fix): CockroachDB could previously, in rare cases, encounter an internal error when evaluating crdb_internal.range_stats builtin (which powers SHOW RANGES command among other things). The bug was introduced in 22.2.0 and is now fixed.

[cockroach-teamcity]

This change is 

[ajwerner]

I don't understand the contracts, and find the need to do this surprising, but I'll happily stamp it.

[DrewKimball]

Huh, I never really thought about this, but I guess we explicitly set the selection vector length everywhere else.

Reviewed 2 of 2 files at r1, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @yuzefovich)

[DrewKimball]

Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @yuzefovich)

---

pkg/col/coldata/batch.go line 51 at r1 (raw file):

	// not been filtered out by a previous step.
	// TODO(yuzefovich): consider ensuring that the length of the returned slice
	// equals the length of the batch.

+1

[yuzefovich]

The contract hasn't really been properly defined. I think in many other places we do set the length on the selection vector mostly in order to get bounds-checks elimination, but IIRC there was a similar bug in another operator at some point, so it's definitely worth being explicit about the contract.

TFTRs!

bors r+

[craig]

Build succeeded:

• Bazel Essential CI (Cockroach)

[knz]

I couldn't reproduce the problem when using demo, so there is no regression test.

There was always 5-10 tests failing in CI for #93644 due to this bug so it was quite easy to reproduce actually. I'll rebase and report if the situation improves.

[knz]

yep, the situation is improved. thanks!